What this project does

auto-ontology is an open source Python library from NVIDIA that adds a semantic, traceable reasoning layer for LLMs and AI agents. Enterprise data is scattered across systems with inconsistent metric definitions, and letting an agent query it directly often produces answers you cannot trust. The library auto-generates an "ontology" layer from your enterprise metadata, exposing context such as where data lives, what it means, whether it is usable, how to use it, where it came from, and what it resembles, through MCP, CLI, and API surfaces for any agent framework, as a supplement on top of existing semantic models.

Its core capability breaks into five steps. Discovery generates the enterprise ontology and picks up semantic models you already have. Grounding gives metrics and entities trustworthy business meaning and reconciles conflicting definitions across systems. Validation checks generated queries against certified definitions and lineage, returning a provenance trail. Exposure serves everything through MCP, CLI, and API. Learning feeds certified corrections and usage back into the ontology so the system improves over time.

What problem it solves

A wrong revenue number is worse than no answer. Agents today can read documents and query databases, but speed without reliability has negative value. auto-ontology targets the reliability bottleneck in agentic analytics: a unified semantic catalog and certified definitions make text-to-SQL answers traceable and verifiable, so each agent no longer interprets business logic its own way. Crucially, when a question cannot be answered from available data, it tells you "this is not answerable" instead of fabricating a plausible result. Like What Is claude-mem and How to Use It: Persistent Cross Session Memory for AI Agents, it is part of the agent infrastructure that makes AI output more reliable and auditable.

Language and tech architecture

The backend is primarily Python, using FastAPI with LangChain for conversation and natural language to SQL. The frontend is Next.js 16 (React 19, TypeScript, Tailwind CSS 4) with Better Auth and Prisma handling the web UI, auth, and API gateway. An ingestion service is a separate FastAPI process that reads tabular data and writes embeddings. Storage uses Postgres with pgvector (PostgreSQL 17) for metadata, catalog, ontology, and vectors, with credentials optionally held in HashiCorp Vault. The MCP server, built on FastMCP, exposes read only agent tools such as check_readiness, search_terms, get_term_sql_attributes, check_answerable, and ask_question. Inference calls NVIDIA NIM, either the hosted integrate.api.nvidia.com or a self hosted NIM. It connects to 10 data source types, including PostgreSQL, MySQL, Databricks, Snowflake, ClickHouse, Trino, Apache Kyuubi, HeavyDB, DuckDB, and SQLite, and is compatible with the Apache Ossie semantic model exchange standard. To build a similar Python plus Docker local dev stack on Windows, see Install and Set Up WSL2 on Windows to Run Linux.

License

Apache-2.0

In 2026, trustworthy and explainable agent driven data analysis has moved from a nice to have to a hard requirement. Backed by NVIDIA, the project attacks text-to-SQL hallucinations at the root with certified ontologies and provenance, hitting the exact pain point of enterprise adoption. On top of that, it exposes capabilities through MCP, CLI, and API, covers 10 data sources, and supports Kubernetes deployment, giving it notably higher engineering maturity. It gained about 2,050 stars on its launch day, leading this batch of new projects.

Project address

GitHub repository https://github.com/NVIDIA/auto-ontology